Bug 83665

Summary:

[TexmapGL] Reduce the number of glTexSubImage2D calls

Product:

WebKit

Reporter:

Igor Trindade Oliveira <igor.oliveira>

Component:

WebCore Misc.

Assignee:

Igor Trindade Oliveira <igor.oliveira>

Status:

RESOLVED FIXED

Severity:

Normal

CC:

dongseong.hwang, noam, webkit.review.bot

Priority:

Version:

528+ (Nightly build)

Hardware:

Unspecified

OS:

Unspecified

Attachments:

Description	Flags
Patch	none
Patch	none

Igor Trindade Oliveira

Reported 2012-04-10 23:18:05 PDT

When the OpenGLES driver does not support GL_EXT_unpack_subimage, BitmapTextureGL::updateContents is transfering the pixels row by row, sometimes calling glTexSubImage2D more than 40 times. This approach is ok for almost all drivers, however for some low end drivers can be a problem.

Attachments
Patch (2.80 KB, patch) 2012-04-10 23:24 PDT, Igor Trindade Oliveira	no flags	Details Formatted Diff Diff
Patch (2.78 KB, patch) 2012-04-11 10:41 PDT, Igor Trindade Oliveira	no flags	Details Formatted Diff Diff
Show Obsolete (1) View All Add attachment proposed patch, testcase, etc.

Igor Trindade Oliveira

Comment 1 2012-04-10 23:24:24 PDT

Created attachment 136630 [details] Patch Proposed patch.

Noam Rosenthal

Comment 2 2012-04-11 05:07:06 PDT

Comment on attachment 136630 [details] Patch Any idea if this is actually faster?

Igor Trindade Oliveira

Comment 3 2012-04-11 10:41:13 PDT

Created attachment 136695 [details] Patch Proposed patch v2.

Igor Trindade Oliveira

Comment 4 2012-04-11 10:47:07 PDT

i did some tests using callgrind and it shows improvements. But it also depends of the graphics driver, in mali drivers i believe the memory is shared between the cpu and gpu so the old code can be fast. (In reply to comment #2) > (From update of attachment 136630 [details]) > Any idea if this is actually faster?

Noam Rosenthal

Comment 5 2012-04-11 11:02:37 PDT

Comment on attachment 136695 [details] Patch This still feels a bit like a micro-optimization that might have different results on different graphics drivers... I'd rather come back to this after we know the pros/cons better.

Dongseong Hwang

Comment 6 2012-06-15 22:24:47 PDT

I think less glTexSubImage2D call is better, although I have no clue to convince you. http://origin-developer.nvidia.com/docs/IO/8230/BatchBatchBatch.pdf?q=docs/IO/8230/BatchBatchBatch.pdf It is be widely known that less drawing call, better performance. So, I think the fact may apply glTexSubImage2D, also, because glTexSubImage2D causes data transfer from cpu to gpu like drawing call.

Noam Rosenthal

Comment 7 2012-06-16 06:57:55 PDT

Comment on attachment 136695 [details] Patch OK, I think you have some merit.

WebKit Review Bot

Comment 8 2012-06-16 09:16:36 PDT

Comment on attachment 136695 [details] Patch Clearing flags on attachment: 136695 Committed r120527: <http://trac.webkit.org/changeset/120527>

WebKit Review Bot

Comment 9 2012-06-16 09:16:40 PDT

All reviewed patches have been landed. Closing bug.

Note You need to log in before you can comment on or make changes to this bug.